Code corrupted silently when import decord before torch

[zhangbw17]

This issue occurs when I import decord before torch, and then place nn.Module on the GPU.

import decord
import torch

torch.nn.Linear(3, 3).cuda()

It corrupted silently using python3 debug.py, and reported Segmentation fault (core dumped) when running in terminal.
Instead, the following code runs well,

import torch
import decord

torch.nn.Linear(3, 3).cuda()

[YinAoXiong]

same problem

[tongda]

same with me. versions:

python: 3.10.14
decord: 0.6.0
pytorch: 2.3.0

and cuda related libs installed by pip:
nvidia-cublas-cu12 12.1.3.1
nvidia-cuda-cupti-cu12 12.1.105
nvidia-cuda-nvrtc-cu12 12.1.105
nvidia-cuda-runtime-cu12 12.1.105
nvidia-cudnn-cu12 8.9.2.26
nvidia-cufft-cu12 11.0.2.54
nvidia-curand-cu12 10.3.2.106
nvidia-cusolver-cu12 11.4.5.107
nvidia-cusparse-cu12 12.1.0.106
nvidia-nccl-cu12 2.20.5
nvidia-nvjitlink-cu12 12.4.127
nvidia-nvtx-cu12 12.1.105

[Leojc]

same problem.

python: 3.10.13
pytorch: 2.3.0
decord: 0.6.0

[losehu]

me too

• python=3.10
• pytorch=1.13.1
• decord==0.6.0

[ellemcfarlane]

Also happens to me, specifically when

1. importing CPU version of decord before torch via import decord (have not tested with GPU version)
2. moving my model to gpu e.g. model.to(torch.device("cuda"), dtype=torch.float16)

Note1: this occurs even when just importing decord but not actually using it to do anything
Note2: it does not occur when moving the model to cpu e.g. model.to(torch.device("cpu"), dtype=torch.float16), so given that I'm using the cpu-version of decord, there might be a connection there, but regardless, this should not happen.

Issue fixed when: simply importing torch before decord

Specific log when running with huggingface accelerate:
subprocess.CalledProcessError: Command '['python3', 'train.py', <placeholder-args>]' died with <Signals.SIGSEGV: 11>.

without accelerate:
train_script.sh: line 3: 2131749 Segmentation fault (core dumped) python3 train.py <placeholder-args>

versions:

python 3.10.14
cuda 12.1
decord 0.6.0
torch 2.4.0
torchvision 0.19.0

[a-r-r-o-w]

Issue fixed when: simply importing torch before decord

Thank you for saving my time @ellemcfarlane! I was stuck on this for quite a bit - very weird/unexpected how this works

[lhoestq]

it also happened to me with duckdb, which needs to be imported before decord or it crashes on import